Comparing strategies to estimate the association of obesity with mortality via a Markov model.
Chen, B. E., Graubard, B. I., Flegal, K. M. and Gail, M. H.
Statistics and Its Interface.
We used a first order discrete Markov model to investigate strategies to obtain unbiased estimates
of the relative mortality hazard for comparing obese with non-obese participants. This hazard ratio
is confounded by the fact that obese participants can be either sick or well, as can non-obese
participants, and participants can migrate over time from their initial classification on obesity
and health status. The parameters of the model were estimated from national survey data and used
to illustrate different analytic approaches. The purpose was to compare analytic approaches and not
to provide an analysis of a particular data set. Under this model, short term health-stratum-specific
estimates are unbiased for estimating the health-stratum-specific instantaneous mortality hazard
ratios from obesity, and, updating information on body mass index and disease status during long
term follow-up reduces bias. For follow-up over 10 or 20 years, exclusion of participants with
pre-existing disease, excluding the first five years of follow-up, and methods of analysis that
ignore health status yield biased estimates of the instantaneous mortality hazard ratios. However,
over 10 or 20 year time periods, long-term average mortality hazard ratios or cumulative mortality
relative risks are a better reflection of the total impact of obesity, including its tendency to
accelerate transitions to sickness under this model, than are instantaneous mortality hazard ratios.
Over these longer time periods, average relative hazard estimates or cumulative mortality relative
risks based on initially well subjects, on initially sick subjects, and on the combined initial
population each provide valuable descriptions of associations of obesity with mortality.
